iiinova — where your digital self keeps evolving

Ben Augustine @bencaugustine.bsky.social · 22/09/2026

On one hand, I don't like hiding most of the model in a custom distribution, but on the other hand, the model runs 228x faster here, so.....

010

Ben Augustine @bencaugustine.bsky.social · 25/06/2026

Yea, OK, mixing can be fine with conditional version of SMR, too, with abundant telemetry data. Runs faster, so you can run more iterations per unit time. Poor mixing in conditional model more of an issue for abundance, survival, recruitment. Negbin obsmod here.

000

Ben Augustine @bencaugustine.bsky.social · 24/06/2026

With abundant telemetry data and patience, you can get acceptable (but still not good) mixing for yearly RSF movement parameters in open population SMR. At least when marginalizing out individual ID. Not sure if it's worth making a version for the conditional model where mixing already is not great

110

Ben Augustine @bencaugustine.bsky.social · 22/06/2026

This point is it the discussion, though

000

Ben Augustine @bencaugustine.bsky.social · 17/05/2026

Cluster SCR model for perfectly aggregated and cohesive groups where you observe individual ID but not group ID. Pretty sure. 😆 github.com/benaug/Group...

040

Ben Augustine @bencaugustine.bsky.social · 06/05/2026

Truth is 40 groups, 34 detected groups (n), 204 individuals, ZTPois group size parameter lambda.P=5

010

Ben Augustine @bencaugustine.bsky.social · 06/05/2026

Dusted off this model for group SCR with unknown group IDs. It is difficult to determine which updates are needed to fully explore the posterior, but currently works well with higher group D and smaller group sizes. But posterior bottlenecks otherwise. Need to see if I can alleviate those.

110

Ben Augustine @bencaugustine.bsky.social · 11/04/2026

I never understand what the bots are up to

100

Ben Augustine @bencaugustine.bsky.social · 10/04/2026

Open population SMR scenario with marking effort only in years 1 and 3 out of 6. 60% of marks stay on until following year, 40% make it to 2nd year (unless animal dies), so marks persist into sighting only years.

010

Ben Augustine @bencaugustine.bsky.social · 27/03/2026

Added 2 interior years with no sampling.

000

Ben Augustine @bencaugustine.bsky.social · 27/03/2026

Messing around with Open SCR + telemetry survival. Here, Method A is live capture to deploy a collar, assumed to stay on for 2 years or until they die (starting simple). Method B is some generic capture recapture method, say hair snares. So hair snare in years 1,3,5, live cap in years 2,4.

110

Ben Augustine @bencaugustine.bsky.social · 20/03/2026

Leveled up this OpenPopSCR model with density covs in year 1 and RSF activity center relocation--runs faster, mixes better. Ran this in 1.5 hours. M=225, 5 years, 144 traps/year, 1156 habitat cells. Effective sample sizes: D.beta1 rsf.beta sigma.move 779.5270 397.6123 368.2466

000

Ben Augustine @bencaugustine.bsky.social · 19/03/2026

Still not great mixing, but finding that open pop SCR dispersal scale parameters mix better if you don't keep the activity centers in the model for augmented individuals are not in the population and just propose a new s trajectory when proposing to turn them on

000

Ben Augustine @bencaugustine.bsky.social · 13/03/2026

The Chandler Clark Jolly Seber parameterization implies Poisson entries as M goes to infinity, but in practice we want to set M as low as we can get away with. This causes the variance in recruits to be less than Poisson variance and it shrinks through time. Here "slack" is M - N.super.

000

Ben Augustine @bencaugustine.bsky.social · 08/03/2026

17 years ago. Today, I have to simulate my own bears.

010

Ben Augustine @bencaugustine.bsky.social · 03/03/2026

Aha! Unless I am missing something obvious, the model allows for 0 cohesion and estimates are biased with less than nominal coverage as simulated cohesion increases. Emett et al. considers perfect cohesion, this paper 0 cohesion. detections of group members is independent of others given AC.

100

Ben Augustine @bencaugustine.bsky.social · 03/03/2026

Just saw this group SMR paper and was wondering how they did it without using the Emmett approach with double data augmentation! I got it wrong in my first attempt. This works because of the "using data twice" approach where you model marked data as marked and then combine with unmarked.

120

Ben Augustine @bencaugustine.bsky.social · 14/02/2026

Modified the Royle-Young supermodel to do sign searches (e.g., scat transects). Lets you retain the continuous det locs instead of snapping to effective detector centroid. Assumes perfect compensatory heterogeneity. Within home range resource selection. Seems like an improvement.

100

Ben Augustine @bencaugustine.bsky.social · 08/02/2026

🤔🧐 Some of the better fits. Used dirichlet-multinomial model for telemetry counts in cells, within cell location informs avail sigma. Random RSF slopes for individuals. Need better covariates or spatial random effects, but seems pretty cool. Daily locs, probably need to thin more.

100

Ben Augustine @bencaugustine.bsky.social · 06/02/2026

Not sure who this Ben is, but they freak me out weekly--what did i agree to, and who is asking me questions!?

010

Ben Augustine @bencaugustine.bsky.social · 04/02/2026

Here, movement is less costly along the road. Or say it is a ridge. Could model anisotropic HRs where the angle of elongation isn't the same for all individuals, say ridge lines, valley bottoms. Can do this with current LCP models, but they introduce unwanted individual heterogeneity in space use/p

000

Ben Augustine @bencaugustine.bsky.social · 04/02/2026

Nothing new here, but inference about landscape connectivity and resource selection could differ a lot depending on how availability is defined. Here, decay is halfnormal, hazard, annular-normal (territorial sps?)

240

Ben Augustine @bencaugustine.bsky.social · 31/01/2026

Row number for road was one off before. So they don't hang out in the road if it is costly to move into it unless there are good resources there to offset the movement cost.

000

Ben Augustine @bencaugustine.bsky.social · 31/01/2026

Simulation of home range placement and within home range resource selection with connectivity constraints for availability. Assuming the road is narrow relative to grid res and animals can spend time in road cells if they contain good habitat, just are less likely to cross them.

120

Ben Augustine @bencaugustine.bsky.social · 20/01/2026

Exploring impact of search effort resolution for Royle-Young model. How accurate does it need to be? In this scenario, probably 1sigma or less. 2sigma only -4% bias, not bad. 4sigma -8%

100

Ben Augustine @bencaugustine.bsky.social · 13/01/2026

One thing regular SCR does not account for is if you have a hard boundary like hab mask, detection probability should generally go up inside the hab mask if the detection function extends beyond it. Ignoring this can introduce false individual heterogeneity in p. Proper RSF model is one solution.

120

Ben Augustine @bencaugustine.bsky.social · 19/11/2025

Recently came across this Dorazio et al. paper showing that parameter expansion can make probit multispecies occupancy models much more efficient to sample. besjournals.onlinelibrary.wiley.com/doi/full/10.... I wanted to learn a new trick so I made a nimble version. github.com/benaug/Probi...

220

Ben Augustine @bencaugustine.bsky.social · 15/10/2025

After all the study planning, field work, model development/ refinement, and computation--Behold! the posterior distribution

150

Ben Augustine @bencaugustine.bsky.social · 11/09/2025

Matern Type III repulsive point processes in NIMBLE. Pretty cool approach and does not require a birth-death process for estimation, so it is fast and mixes well. Can add SCR observation model. github.com/benaug/Mater...

020

Ben Augustine @bencaugustine.bsky.social · 26/08/2025

Simulate a point process, simulate uniform(0,1) ages for each point, put points in age order, then thin using a half-normal kernel.

100

Ben Augustine @bencaugustine.bsky.social · 26/08/2025

Got generalized Matern Type III thinning process for point repulsion working using data augmentation. Cool approach, but pretty slow.

100

Ben Augustine @bencaugustine.bsky.social · 13/08/2025

Figured out how this data augmentation approach to is related to that presented by Schofield and Barker (2014). Theirs can be specified with bugs code, which is very slick! As coded, mine runs 2x as fast with better ESS/iteration. Cool stuff. More formal details soon. github.com/benaug/SCR-N...

010

Ben Augustine @bencaugustine.bsky.social · 02/08/2025

Setting up final run for SMRing ground squirrels. Mb for marking process, negative binomial for sighting process, both sex-specific.

110

Ben Augustine @bencaugustine.bsky.social · 12/07/2025

The overdispersion is strong in the SMR data set

000

Ben Augustine @bencaugustine.bsky.social · 20/06/2025

Marginal version of SMR with interspersed marking and sighting and all sample types is pretty slick.

010

Ben Augustine @bencaugustine.bsky.social · 20/05/2025

Top 5 things I want to write up from my Github repositories. Someone hire me for a while and we can pump these out. Will describe some more later.

010

Ben Augustine @bencaugustine.bsky.social · 24/04/2025

Works with within home range resource selection, too. Here, availability distribution is BVN, use distribution is normalized product of availability distribution and RSF.

000

Ben Augustine @bencaugustine.bsky.social · 24/04/2025

This looks promising approach for repulsion between individuals based on space use overlap instead of distance. Generalized Matern Type III Process thins a Poisson point process as a function of space use overlap, where space use is BVN.

210

Ben Augustine @bencaugustine.bsky.social · 21/04/2025

Precognitive subskeeting? Damn, he's good.

120

Ben Augustine @bencaugustine.bsky.social · 18/04/2025

Pop simulator with habitat and density effects on reproduction, survival, and movement. Females only. Grey indicates 95% HR. Survival: red dots adult deaths, gold, 1 or more dependent deaths. Movement, red is adult HR relocation, gold subadult dispersal. Reproduction: gold indicates a new litter.

010

Ben Augustine @bencaugustine.bsky.social · 09/04/2025

The fastest way to Mb is to create first and subsequent capture histories and exposure matrices. Reduces data dimensions to equivalent of 2 occasions. Added this to my SCR Dcov repo. github.com/benaug/SCR-D...

100

Ben Augustine @bencaugustine.bsky.social · 08/04/2025

Last single-catch thing for now. Using Mb here, trap-happy. Obsmod parameters weakly identifiable esp. first cap p0 and sigma. What happens is trap happy inds almost always get to traps first and mostly block first capture events.

100

Ben Augustine @bencaugustine.bsky.social · 07/04/2025

The main problem with single catch traps is that you block more captures in high D areas than low D areas and underestimate Dcov betas using regular SCR. Can see that effect here.

010

Ben Augustine @bencaugustine.bsky.social · 06/04/2025

Latent ID SCR approach to Efford 2004 Oikos single catch traps modeling observed captures conditioned on the captures that would have happened if traps did not fill up or prevent captured individuals from being captured at other traps on the same occasion. github.com/benaug/SCR_S...

000